Anthropic Launches Browser-Based Claude AI: Navigating the Future of Safe AI Integration
A groundbreaking research preview of an AI agent, powered by Claude AI models from Anthropic, has been unveiled. The AI agent, named Claude for Chrome, will be accessible to a select group of 1,000 subscribers on the company’s Max plan, priced between $100 and $200 monthly. Interested users can also join a waitlist.
By installing an extension in Google Chrome, users can now interact with Claude within a sidebar window that maintains browser context. The AI agent is also granted permission to perform tasks on behalf of users within the browser.
The race for AI-integrated browsers is heating up, as leading labs aim to provide more seamless connections between users and AI systems via browser integrations. Recent entries include Comet by Perplexity, featuring an AI agent capable of offloading tasks for users, and reports suggest that OpenAI is preparing to launch a similar product. Google’s Chrome has recently seen the addition of Gemini integrations.
The development of AI-powered browsers takes on increased urgency in light of potential regulatory action against Google, with a ruling imminent in its antitrust case. There are rumors that the court may require Google to sell its Chrome browser. Perplexity has reportedly offered $34.5 billion for Chrome, and OpenAI’s CEO Sam Altman has expressed interest in making a similar purchase.
In a blog post announcing the research preview, Anthropic acknowledged potential safety risks associated with AI agents that have browser access. Last week, Brave’s security team identified vulnerabilities in Comet’s browser agent that could make it susceptible to prompt-injection attacks. Perplexity has reportedly addressed this issue.
Anthropic aims to use this research phase to identify and mitigate novel safety risks. The company has already implemented several defenses against prompt injection attacks, reducing their success rate from 23.6% to 11.2%. Users can restrict Claude’s browser access to certain sites through the app’s settings, and by default, Claude is blocked from accessing websites offering financial services, adult content, or pirated content. Additionally, the AI agent will seek user approval before performing high-risk actions such as publishing, purchasing, or sharing personal data.
Anthropic has previously developed an AI model capable of controlling a PC, launched in October 2024, though testing at the time revealed the model to be slow and unreliable.
Modern browser-based AI agents, like Comet and ChatGPT Agent, have demonstrated improved reliability when it comes to offloading simple tasks for users. However, these agentic systems continue to grapple with more complex problems.
